Does naming the platform change how AI agents judge sponsored listings?

In matched hotel-listing choice experiments, an AI agent told it was delegated by a traveler penalized a sponsored listing — it chose the listing 50.2 percentage points less often than a similar organic listing. When the same system prompt instead named a booking platform as the agent's employer ("platform delegation"), the penalty attenuated to 29.2 percentage points. The paper replicates this pattern "across several foundation models, reasoning depths, and system prompt variations," and reports it also shows up in the agent's own reasoning traces: platform delegation "systematically reduces the skepticism that sponsorship disclosures trigger."

The paper frames this through fiduciary language: naming the platform or the consumer as the agent's "principal" in a system prompt sets up what it calls "sponsorship bias" — a more favorable evaluation of sponsored listings when the platform, rather than the consumer, is named as the agent's employer. The mechanism it gives is context-dependent role adoption, in which an LLM "tailors its outputs to the inferred perspective" of whichever party is named, drawing on "the widespread skepticism shoppers express toward paid placements" absorbed from training text, then discounting that skepticism when the named principal is a party that profits from the placement rather than one exposed to its risk. A second study found the divergence between the two delegates "widens when the paid placement is attributed to the platform," and that explicit wording ("Sponsored" instead of "Promoted") lowers choice of paid listings but "does not close this gap when the platform is named."

The paper names its diagnosis in the same terms Can traditional governance tools actually control AI agents? uses for AI-agent delegation generally: a conflict of duty between two principals. Where that note's source argues conventional monitoring and incentive tools fail because agent reasoning is opaque, this paper tests a narrower version of the same conflict experimentally and finds a sentence-level fix, detailed in the companion note Can naming the consumer fix AI agent sponsorship bias?. It also sharpens Can EU consumer law protect AI agents as algorithmic buyers?: that article argues sponsorship-disclosure law assumes a human reader processes a "Sponsored" label with skepticism, and this paper shows the disclosure now reaches the AI agent instead, where the same label triggers more or less skepticism depending on who the agent is told it works for.

Both studies were run in a single domain, hotel booking, and the authors flag that whether this role-induced bias extends to other advisory domains, such as financial advice or healthcare, is untested. The paper also does not establish what discount a sponsored listing epistemically deserves, so it cannot say platform delegation produces an objectively wrong evaluation — only that it moves the evaluation further toward the platform's interest than either consumer delegation or the model's own role-neutral default produced. The implication it draws plainly: sponsorship-disclosure law written for a human reader cannot police this bias on its own, because the entity the disclosure must now persuade is the AI agent itself.